Experimental protocols are commonly communicated as natural-language documents whose ambiguity, omitted assumptions, and inconsistent terminology limit reproducibility and automation. This project will develop domain-specific translation models for converting natural-language laboratory protocols into Culsma, a structured language for representing experimental operations, materials, conditions, state transitions, and execution constraints. The models will be trained on non-sensitive, openly licensed laboratory-protocol corpora and compiler-validated Culsma training pairs, with an initial focus on systems-immunology workflows. We will investigate supervised fine-tuning, synthetic data augmentation, long-context modelling, constrained decoding, and compiler-guided feedback across several model sizes and training configurations. Translation quality will be evaluated not only with conventional language metrics, but also through Culsma parser and compiler acceptance, structural and semantic agreement with reference programs, material-state continuity, and execution-grounded consistency checks. The project will establish whether neural translation combined with formal validation can reliably transform informal experimental instructions into executable and machine-verifiable protocol representations. The existing open-source Culsma implementation provides the parser, compiler, runtime and automated tests needed for this validation loop. The resulting models, evaluation datasets, and validation methods will support reproducible protocol authoring, automated protocol checking, and future integration with laboratory instruments and robotic platforms. No sensitive personal data will be processed on the requested resources. The Culsma implementation, documentation, release records, paper and reproducible examples are available through https://github.com/culsma/culsma.